Method and apparatus for constructing speech decoding network in digital speech recognition

Abstract

The invention discloses a method and an apparatus for constructing a speech decoding network in digital speech recognition. The method includes the following steps: acquiring training data obtained from digital speech recording, the training data including a plurality of speech sections; extracting acoustic features from the training data so as to obtain a feature sequence which corresponding with each speech section; based on the feature sequences and phonemes corresponding with digits in the training data, conducting progressive training beginning with a single phoneme acoustic model to obtain an acoustic model; acquiring a language model, constructing a speech decoding network through the language model and the acoustic model obtained in the training, the language model being obtained by modeling matching relationship of digits in the training data. According to the invention, the method and the apparatus can effectively increase recognition accuracy of digital speech.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a method and device for constructing a speech decoding network in digital speech recognition. Background technique [0002] In digital speech recognition, the existing speech decoding networks can be divided into two categories: one is to use isolated word recognition technology to recognize numbers in speech; the other is to use general continuous speech recognition technology to process numbers in speech. identify. [0003] In digital speech recognition based on isolated word recognition technology, it is required to input digital speech with a clear interval between numbers. If it is continuous digital input, it may lead to unrecognized or incorrect recognition, which greatly reduces the recognition accuracy of digital speech. , digital speech recognition based on isolated word recognition technology has obvious limitations. [0004] Thus, general continuous speec...

AI Technical Summary

Problems solved by technology

However, the recognition object of this technology includes not only numbers, but also other language content, which makes the acoustic model and language model used in this technology too complex, the recognition speed is relatively slow, and it is easy to misrecognize numbers as other polyphonic characters , so that the recognition accuracy of digital speech is not high enough

Even if the recognition objects of the language model in this technology are limited to ten numbers from 0 to 9, the improvement of recognition accuracy is still limited

[0005] It can be seen that the speech decoding network built for digital speech recognition still has the problem of low recognition accuracy

Images